KNEE INJURY DETECTION USING MRI WITH EFFICIENTLY LAYERED NETWORK - PowerPoint PPT Presentation

KNEE INJURY DETECTION USING MRI WITH EFFICIENTLY LAYERED NETWORK (ELNET) Chen-Han Han Tsai, i, Nahum um Kiryati ati, , Eli i Konen, , Iris Eshed, , Arnald naldo o Mayer er

PROBLEM MOTIVATION CONVEN ENTIONA IONAL L KNEE EE EXAMS MS MRI Acquisition Doctor’s Analysis Final Assessment Exam added to Queue Sorted by Exam Date MSK radiologists face a rising work demand each day  Triage improves efficiency by prioritization  Severe cases prioritized first 

PROBLEM MOTIVATION TRIAGE GED D KNEE EE EXAMIN INATIONS IONS MRI Acquisition Doctor’s Analysis Final Assessment Sorted by Level of Severity MSK radiologists face a rising work demand each day  Triage improves efficiency by prioritization  Severe cases prioritized first 

ELNET ARCHITECTURE Fig-1: Illustration and configuration of ELNet.

ELNET CORE COMPONENTS Fig-2: Block with 2 repeats Fig-4: BlurPool Down-sampling Fig-3: Multi-slice Normalization for 3D Inputs

EVALUATION DATASETS  MRNet et Datase taset 1  1370 knee MRI exams *  Labels : ACL tear / Meniscus tear / Abnormalities  Axial, coronal, and sagittal scans provided Axial Plane Coronal Plane  KneeM eMRI I Datas aset et 2  917 knee MRI exams **  Labels: ACL Injured  Sagittal scan provided Sagittal Plane 1 Bien et al, Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet, PLOS Medicine (2018) 2 Štajduhar et al, Semi -automated detection of anterior cruciate ligament injury from MRI, Computer Methods and Programs in Biomedicine (2017)

ELNET SETUP Detect ection ion Object ectiv ive Multi-Sl Slic ice e Norm Image e Modalit lity K Number er of Paramet eter ers Meniscus Tear Contrast Norm Coronal 4 ~ 0.2 M (850 kB) ACL Tear Layer Norm Axial 4 - Abnormalities Layer Norm Axial 4 - ACL Tear (KneeMRI) Contrast Norm Sagittal 2 ~ 0.05 M (438 kB)  ELNet is trained from scratch  Previous SOTA MRNet ~183M parameters for each objective

MRNET EVALUATION Fig-5: Evaluation of ELNet and MRNet performance on the MRNet Dataset

KNEEMRI EVALUATION Fig-6: Comparison of ELNet performance across all 5 folds on the KneeMRI dataset

KNEEMRI EVALUATION Fig-7 : ROC’s of ELNet of KneeMRI Dataset across 5 folds

MODEL INTERPRETATION Fig-8: Full-Grad visualization highlighting the tear locations in the knee

SUMMARY  ELNet features  Lightweight  Adequate performance  Easily trained from scratch  May be applied to other pathologies involving 3D images (MRI, CT, etc.)